Anti-Collusion Levers Are Not Free: Measuring What Monitoring Costs LLM Markets

The question behind this study is straightforward: which intervention can reduce collusion between LLM pricing agents at the lowest cost to the market? Previous work has mainly compared interventions by how much collusion they prevent, without measuring what those interventions cost when there is no collusion to prevent. We use the continuous double auction from Agrawal et al. (2025) to investigate this trade-off. First, we audited the authors' public runs and found that the LLM monitor triggered in 20 of 20 oversight runs, within seven rounds, restricting sellers to five-character messages for 77 to 97% of each run. In practice, this makes oversight behave more like a delayed communication ban than light-touch monitoring. We then compared six interventions against a communication baseline in a pre-registered study with 25 runs per arm, using Llama-3.3-70B sellers. Neither Llama-3.3-70B nor GPT-4.1-mini sellers colluded when communication was allowed, meaning there was no collusion for the interventions to reduce. Instead, we measured their costs in a non-colluding market. The monitor punished innocent markets in up to 48% of runs, the monitoring notice reduced buyer welfare by 22%, and banning communication cost 13% of trades. In contrast, adding one no-regret competitor increased buyer welfare by 26%. A further 75 pre-registered runs confirmed both welfare effects, although the notice's effect on trade volume did not replicate. These findings show why the costs of false positives must be measured alongside the benefits of anti-collusion interventions.